【作者】 汪成龙；

【导师】 李小昱；

【作者基本信息】 华中农业大学 ， 农业电气化与自动化， 2014， 博士

【摘要】 马铃薯内外部品质的检测直接关系到其加工利用率和增值率,同时也是马铃薯工业化生产加工的首要步骤。近年来,虽然机器视觉和近红外光谱技术分别在马铃薯外部和内部品质检测和分级研究中取得了一定的进展,但还存在无法同时对马铃薯内外部品质同时进行检测的问题。针对这一问题,该文利用机器视觉技术和近红外光谱技术,研究了基于多源信息融合技术的马铃薯分级无损检测方法。试验以克新一号马铃薯为研究对象,对畸形、黑心、机械损伤、发芽和合格等5类不同内外部品质的样本进行分级检测研究,研究了多源信息融合技术检测马铃薯品质的图像和光谱特征提取方法以及融合方法,并最终建立了马铃薯分级融合模型。1)为有效的避免背景对马铃薯图像分割的干扰,该文提出了视觉显著性与色调维相结合的Saliency-H分割方法,并比较了其与灰度分割法和色调维分割法的分割效果。灰度分割法由于其无法分割出完整的马铃薯区域,故不适用于在线马铃薯图像分割,而色调维分割法和Saliency-H维分割法均能完整分割出马铃薯区域,其中Saliency-H维分割法在分割速度、数据压缩和马铃薯定位等方面较色调维分割法具有较大的优势,色调维分割法平均每幅图像耗时为551.7ms,而Saliency-H法减少了74ms,仅需477.7ms。2)针对马铃薯表面灰度不均匀,图像特征难以有效覆盖马铃薯样本集的问题,该文提出了灰度梯度与流形学习组合的方式提取马铃薯图像特征,比较了不同的图像特征组合方式对模型的影响。文中所采用的灰度梯度算法为Freeman链码和方向梯度直方图,流形学习算法为等距映射和主成分分析,在4种算法组合中,方向梯度直方图与主成分分析为最优组合,建模所需图像特征数量最少,仅需23维特征(10维方向梯度直方图特征与13维主成分特征),模型即可达到最优。3)建立了基于机器视觉技术的马铃薯分级检测模型。图像灰度梯度与流形学习特征不同的组合方式所建4个模型对马铃薯外部品质(畸形、机械损伤和发芽)的分级能力均高于内部品质(黑心),其中方向梯度直方图与主成分特征组合而成的图像特征所建模型最优,对畸形、机械损伤和发芽样本的识别率分别为93.75%、83.33%和95.45%,而对黑心和合格样本的识别率分别仅为77.27%和71.43%。4)建立了基于LabVIEW平台的马铃薯外部品质在线检测系统。以38个不同外部品质的马铃薯样本为检测对象,对长径、短径、高径、薯形(类圆、椭圆、长形)、畸形、机械损伤、发芽和合格8项外部品质指标进行检测,畸形、机械损伤、发芽和合格4项外部品质定性指标的识别率为89.47%,对类圆、椭圆和长形3类马铃薯的识别率为100%,对长径、短径和高径的检测最大误差分别为2.9mm,2.0mm和1.0mm,单幅图像平均耗时为100ms。结果表明该文提出的马铃薯图像分割算法、特征提取方法和模式识别方法可实现马铃薯外部品质多项指标的在线检测。5)比较了波段优选算法和流形学习算法的近红外光谱特征提取方法的优劣。文中所采用的波段优选算法为遗传算法和连续投影算法,流形学习算法为拉普拉斯特征映射法、核主成分分析和主成分分析,对于5种近红外光谱特征提取方法所建的马铃薯分级模型,利用主成分分析提取的近红外光谱特征所建模型最优,其最优预处理方法为MSC,最优主成分数量为20,模型对训练集的识别率为97.88%,对测试集的识别率为83.87%,结果表明对于马铃薯近红外光谱特征提取方法,流形学习算法优于波段优选算法,为一个近红外光谱马铃薯分级模型对马铃薯内部品质多项指标的同时检测提供了技术支持。6)建立了基于近红外光谱技术的马铃薯分级检测模型,波段优选算法和流形学习算法所建的分级模型对马铃薯内部品质(黑心和发芽)的识别率均高于外部品质(畸形和机械损伤),其中主成分特征所建模型最优,对黑心和发芽2类样本识别率较高分别为90.91%和95.45%,而对畸形、机械损伤和合格样本的识别率分别仅为75.00%、75.00%和76.19%。7)利用LabVIEW实现了近红外光谱技术的马铃薯分级检测系统软件,黑心、发芽和合格3项马铃薯内部品质的识别率达到95.45%。在算法执行效率方面,平均每条光谱的预处理耗时为3.4ms,20维主成分提取耗时为14.6ms,建模耗时5137ms,利用模型对单条光谱测试,平均耗时为15.0ms,可实现30条/s的检测效率,为一个近红外光谱模型在线检测马铃薯内部品质多项指标提供了技术支持。8)确定了多源信息融合技术检测马铃薯内外部品质的融合方法,比较了不同融合方法所建马铃薯分级检测模型的检测精度。以畸形、机械损伤、黑心和发芽和合格5类马铃薯样本为研究对象,建立马铃薯内外部品质多项指标的多源信息融合模型,决策层融合方面,采用机器视觉和近红外光谱所建支持向量机模型的概率输出为基本概率赋值函数,以DS证据理论为决策层融合方法,建立决策层融合模型,对训练集的识别率为100.00%,对测试集的识别率为93.55%；特征层融合方面,利用方向梯度直方图与主成分分析组合的方式提取图像特征,利用主成分分析提取光谱特征,将图像和光谱特征作为模式识别的输入,分别利用Adaboost和支持向量机建立特征层融合模型。AdaBoost所建模型对训练集的识别率为100.00%,对测试集的识别率为91.40%,支持向量机所建模型对训练集的识别率为100.00%,对测试集的识别率为95.70%。结果表明对于马铃薯内外部品质多项指标的检测,支持向量机特征层融合优于DS决策层融合,DS决策层融合优于AdaBoost特征层融合,那么,支持向量机特征层融合模型为最优的马铃薯分级融合模型。9)建立了基于多源信息融合技术的马铃薯分级检测模型,可实现一个融合模型同时检测马铃薯内外部品质多项指标。融合模型对畸形、黑心、机械损伤、发芽和合格样本识别率分别为100.00%、95.45%、91.67%、100.00%和90.48%,相对于机器视觉所建马铃薯分级检测模型,融合模型对畸形、机械损伤、黑心和发芽和合格5类马铃薯样本的识别率分别提高了6.25%、18.18%、8.34%、4.55%、19.05%,而对于近红外光谱所建马铃薯分级检测模型,融合模型对上述5类马铃薯样本的识别率则分别提高了25.00%、4.54%、16.67%、4.55%、14.29%。10)利用LabVIEW实现了多源信息融合技术的马铃薯分级检测模型,并对图像分割、图像特征提取、光谱预处理、光谱特征提取、相应指标测取等进行了测试,每个样本的平均总耗时低于140ms,能实现每秒7组图像和近红外光谱数据的处理速度。结果表明,基于多源信息融合的马铃薯分级检测模型的识别率优于单一的机器视觉或近红外光谱所建模型,为利用多源信息融合技术在线检测马铃薯内外部品质多项指标提供了技术支持。更多还原

【Abstract】 Detection of potato external quality is directly related to the processing and utilization rate and growth rate, and the commercial production and processing of potato is the first step. In recent years, although the machine vision and near infrared spectroscopy has respectively made certain progress in the study of detection and classification for external and internal quality of potato, but the detection is not at the same time. Aiming at this problem, the paper uses the technology of machine vision and near infrared spectroscopy, studied the nondestructive detection method of potato based on multi-source information fusion technique. Using Ke Xin Yi Hao potatoes as the research object,5kinds of sample set include deformity, black heart, mechanical damage, eye and normal. Characteristics extraction method of potato images and spectra,and the fusion method for potato detection, are studied, at the same time the potato fusion classification model is established.1) To avoid the interference of background on potato image segmentation, the paper proposed a segmentation method based on visual saliency and the hue dimension, and compared it with the gray segmentation method and the hue dimension segmentation. The gray segmentation method due to its inability complete of potato area, is not suitable for online potato image segmentation. While the hue dimension segmentation and Saliency-H dimension segmentation method can segment the complete potato area, among them Saliency-H dimension segmentation method has great advantage in the segmentation speed, data compression and potato positioning. The simulation on Matlab platform shows, the average time-consuming for hue dimension segmentation method is551.7ms, Saliency-H method reduce74ms, only needs477.7ms.2) Because of the uneven gray of potato surface, image features are difficult to be effectively covered potato samples, the paper proposed the gray gradient and manifold learning combination way for extracting potato image features, models which build are compared. The gray gradient algorithm is adopted in this paper for the Freeman chain code and histogram of oriented gradients, isometric mapping and principal component analysis for manifold learning algorithms. Histogram of oriented gradients and principal component analysis is the best combination of image features, which modeling required only23dimensional feature (feature10dimensional direction the gradient histogram features and13dimensional principal component) to achieve the optimal model.3) Potato grading and detection model based on machine vision is established. The gray gradient and manifold learning characteristics in different combinations for potato external quality (deformity, mechanical damage and eye) classification capacity is higher than the internal quality (black heart), in which histogram of oriented gradients and principal component features is the best combination for modeling. The recognition rates of deformity, mechanical injury and eye sample were93.75%,83.33%and95.45%, and the recognition rates of black heart and the normal sample were only77.27%and71.43%.4)A potato external quality online detection system is builded based on LabVIEW platform. After detect38potato samples with different external quality, identification rate of deformity, mechanical damage, eye and qualified samples was89.47%. Tuber shape is100%, and the long diameter, short diameter and height of the maximum error are respectively2.9mm,2.0mm and1.0mm. Average time for a single image100ms. Results show that the proposed potato image segmentation, feature extraction and pattern recognition methods can realize online detection of potato external quality indicators.5) Effects of band selection algorithms and manifold learning algorithms on potato grading model were compared. Band selection algorithms in the paper include genetic algorithm and successive projections algorithm, and manifold learning algorithms include Laplasse feature mapping method, kernel principal component analysis and principal component. Potato grading model using the principal component analysis of spectral feature is the optimal. The optimal pretreatment method is MSC, and the best number of principal components is20. The recognition rate for training set is97.88%, and83.87%for test set. The results shows that for potato spectral feature extraction, manifold learning algorithm is better than the band selection algorithm.6) Potato grading and detection model based on near infrared spectrum technique is established. Either Band selection algorithm or manifold learning algorithm, the recognition rate of internal quality for potato (black heart and eye) are higher than the external quality (deformity and mechanical damage), the principal component feature model is the optimal, in which the black heart and eye2types of samples has high recognition rates90.91%and95.45%, while the deformity, mechanical damage and the normal sample were only75%,75%and76.19%.7) Use Lab VIEW to establish the potato internal quality detection model based on near infrared spectroscopy. For black heart, eye and qualified potatoes, the identification rate of near infrared reflectance spectroscopy was95.45%. For algorithm execution efficiency, time-consuming of each pretreatment of near infrared reflectance spectroscopy is3.4ms,20dimensional principal component extraction time is14.6ms, the modeling time5137ms. Using the model test of single spectra, average time was15.0ms, the detection efficiency can achieve30/s, which provides technical support for only using a near-infrared spectroscopy model online detection of potato internal quality indicators.8) The fusion method for the detection of potato external quality is determined, and effects of different fusion methods on modeling results for potato grading and detection are compared. In decision level fusion, the probability outputs of machine vision and near infrared spectra based on support vector are gained. Using DS evidence theory method for decision level fusion, decision fusion model is established. Recognition rate of training set is100%, for test set is93.55%. In feature level fusion, the combination of histogram of oriented gradients and principal component analysis is the image features extraction method, and the principal component is used for spectral feature extraction, using image and spectral features as the inputs of pattern recognition, feature level fusion models respectively established by Adaboost and support vector machine. For AdaBoost model, recognition rate of training set is100%, of test set is91.40%. For support vector machine model, recognition rate of training set is100%, of test set is95.70%. Results showed that the detection for potato’s several quality indexes, support vector machine feature fusion and is better than DS decision fusion, and DS decision fusion outperforms AdaBoost feature level fusion, feature level fusion.9) Potato grading and detection model based on multi-information fusion technology is established, which can detect a number of quality indicators of potato. Recognition rates of DS evidence theory model for deformity, black heart, mechanical damage, eye and normal sample are respectively100%,95.45%,83.33%,100%and85.71%. Recognition rates of Adaboost model of deformity, black heart, mechanical damage, eye and normal sample are respectively100%,95.45%,83.33%,100%and76.19%. Recognition rates of support vector machine model for deformity, black heart, mechanical damage, eye and normal sample are respectively100%,95.45%,91.67%,100%and90.48%. Results showed that the fusion model identification for potato were superior to machine vision and near infrared spectra, the support vector machine model is the best, recognition rates of the deformity, black heart, mechanical damage, eye and normal samples reaches more than90%, which providing technical support for multi-source information fusion technology online detection of a number of potato quality indicators.10) Using LabVIEW For the testing of image segmentation, feature extraction, image preprocessing, feature extraction, average total time of each sample is less than140ms, can achieve7per second speed of processing image and near infrared spectral data. The results show that, multi-source information fusion identification model is better than machine vision and near infrared spectroscopy, which provides the technical support for potato external quality indicators online detection by using multi-source information fusion technology.更多还原